Using AVL Data to Measure the Impact of Traffic Congestion on Bus Passenger and Operating Cost

Transcription

1 Using AVL Data to Measure the Impact of Traffic Congestion on Bus Passenger and Operating Cost A Thesis Presented By Ahmed Talat M. Halawani to The Department of Civil and Environmental Engineering in partial fulfillment of the requirements for the degree of Master of Science in Civil Engineering in the field of Transportation Engineering Northeastern University Boston, Massachusetts December, 2014

2 ii ABSTRACT Letting buses operate in mixed traffic is the least costly way to accommodate transit, but that exposes transit to traffic congestion which causes delay and service unreliability. Understanding the real cost that traffic congestion imposes on both passengers and operating agencies is critical for the efficient and equitable management of road space. This study aims to develop a systematic methodology to estimate those costs using Automated Vehicle Location data. Traffic congestion increases cost to both transit operators and passengers. For transit operators, congestion results in longer running times and increased recovery time. To passengers, traffic congestion increases riding time and, because of how congestion increases unreliability, waiting time. Using data from a low-traffic period as a baseline, incremental running time in each period can be calculated. However, some of this incremental running time is due to the greater passenger volumes that typically accompany higher traffic periods. Passenger counts and a regression model for dwell time, estimated from detailed ride check data, are used to estimate the passenger volume effect on running time so that incremental delay due to congestion can be identified. Cost impacts for operators and passengers follow directly. Observed running time variability is a combination of variability due to greater demand, variability in the schedule, inherent variability in running time, variability due to imperfect operating control, and variability due to traffic congestion. Methods are developed to estimate the first four components so that incremental variability due to traffic congestion can be identified for each period, again using a low traffic period as a baseline. From this incremental variability, we can estimate the additional recovery time needed as well as increases in passenger waiting time and potential travel time, which the difference between budgeted travel time and actual travel time.

3 iii The methodology was tested on nine different bus routes including both high and low frequency routes. Overall, the average impact on operating cost is $20.4 per vehiclehour, and the average impact to passengers is $1.30 per passenger; naturally, these impacts are far greater during peak periods.

4 iv ACKNOWLEDGMENTS First and foremost I would like to express my special appreciation and thanks to my advisor, Prof. Peter G. Furth, who offered his continuous advice and encouragement through the past two years. I have been extremely lucky to have a supervisor who has a great personality, wisdom and knowledge. I would also like to thank my committee member, Prof. Haris N. Koutsopoulos, for his advice. I am grateful to Dr. Daniel Dulaski for his instruction during my study at Northeastern University. This paper would not have been completed without the willingness and support of Melissa Dullea, Samuel Hickey, and David Schmeer at MBTA who provided us with all the needed data for this thesis. I also thank the MIT transit research group for providing us with a sample of the AVL data that we used as a first step in exploring the data. I would also like to thank my parents and brothers who were always supporting me and encouraging me with their best wishes. Last but not least, I would like to thank my wife and best friend, Alyaa Alharbi, for her love, patience, and understanding. Finally, I would like to thank my daughter, Basema, who has been such a great inspiration to me.

5 v TABLE OF CONTENTS ABSTRACT... ii TABLE OF CONTENTS... v LIST OF TABLES... vii LIST OF FIGURES... ix Chapter 1. Introduction Overview Research Objective Thesis Organization... 3 Chapter 2. Literature review Measuring Congestion AVL systems Reliability Conclusion... 8 Chapter 3. Data Sources Automated Vehicle Location system (AVL) Heartbeat Data Time-point Data Announcement Record Data Automated Passenger Counting Data (APC) Chapter 4. AVL Data Analysis Methodology Announcement record data processing Time-point data processing Evaluation & Suggestion for the Reviewed AVL Archived Data Chapter 5. Methodology Stop Time Model Dwell Time Model Lost time Grouping trips Average lower speed impact Variability in Running Time impact... 26

6 vi Variability at the trip level Variations from the scheduled running time V FSch (RT) Adjust running time variation for greater demand Impact on Operating Cost Variability at stop level Impact on waiting time with high frequency service Impact on waiting time with low frequency Service Impact on potential (Budgeted) Travel Time Value of time Summary AVL-Free Methodology The Number of Stops Made Model Summary Application to MBTA Route Annual Impact Analyzing route 1 using scheduled RT Chapter 6. Results Chapter 7. Summary and Conclusions Conclusion Future Research REFERENCES APPENDIX A APPENDIX B Route Route 66 (66_6) Route Route Route 39 (39_3) Route 99 (99_7) Route Route

7 vii LIST OF TABLES Table 1 Description of Heartbeat Data Table 2 Description of Time-point Data Table 3 the Components of Announcement record Data Table 4 lost time components and their values Table 5 data grouping periods Table 6 Annual cost component summery for a period p Table 7 trips throughout a year for Route 1, MBTA Table 8 the data sources and its outputs Table 9 Route 1 variables and corresponding adjustments (using AVL data) Table 10 Traffic Congestion Impact on Route 1, MBTA, Boston Table 11 Route 1 variables and corresponding adjustments (using scheduled RT) Table 12 Traffic Congestion Impact on Route 1(using scheduled running time data) Table 13 list of the chosen routes Table 14 Summary Annual Impact of Traffic Congestion on the chosen routes Table 15 Traffic congestion impact per Veh-hr Table 16 Abbreviations Table 17 Route 23 variables and corresponding adjustments (using AVL data) Table 18 Traffic Congestion Impact on Route 23 (Using AVL Data) Table 19 Route 23 variables and corresponding adjustments ( using scheduled running time) Table 20 Annual Impact of The Traffic Congestion Impact on Route 23 (Using Scheduled Running Time Data) Table 21 Route 66 variables and corresponding adjustments (using AVL data) Table 22 Traffic Congestion Impact on Route 66 (Using AVL Data) Table 23 Route 23 variables and corresponding adjustments ( using scheduled running time) Table 24 Traffic Congestion Impact on Route 66(using scheduled running time data) Table 25 Route 77 variables and corresponding adjustments (using AVL data) Table 26 Traffic Congestion Impact on Route 77 (Using AVL Data) Table 27 Route 77 variables and corresponding adjustments ( using scheduled running time) Table 28 Traffic Congestion Impact on Route 77(using scheduled running time data) Table 29 Route 28 variables and corresponding adjustments (using AVL data) Table 30 Traffic Congestion Impact on Route 28 (Using AVL Data) Table 31 Route 28 variables and corresponding adjustments ( using scheduled running time) Table 32 Traffic Congestion Impact on Route 28(using scheduled running time data) Table 33 Route 39 variables and corresponding adjustments (using AVL data) Table 34 Traffic Congestion Impact on Route 39 (Using AVL Data) Table 35 Route 39 variables and corresponding adjustments ( using scheduled running time)... 75

8 Table 36 Traffic Congestion Impact on Route 39(using scheduled running time data) Table 37 Route 99 variables and corresponding adjustments (using AVL data) Table 38 Traffic Congestion Impact on Route 99 (Using AVL Data) Table 39 Route 99 variables and corresponding adjustments ( using scheduled running time) Table 40 Traffic Congestion Impact on Route 99(using scheduled running time data) Table 41 Route 9 variables and corresponding adjustments (using AVL data) Table 42 Traffic Congestion Impact on Route 9, MBTA, Boston Table 43 Route 9 variables and corresponding adjustments ( using scheduled running time) Table 44 Traffic Congestion Impact on Route 9(using scheduled running time data) Table 45 Route 89_ variables and corresponding adjustments (using AVL data) Table 46 Traffic Congestion Impact on Route 89_, MBTA, Boston Table 47 Route 89 variables and corresponding adjustments ( using scheduled running time) Table 48 Traffic Congestion Impact on Route 89(using scheduled running time data) Table 49 Route 89_2 variables and corresponding adjustments (using AVL data) Table 50 Traffic Congestion Impact on Route 89_2, MBTA, Boston Table 51 Route 89.2 variables and corresponding adjustments ( using scheduled running time) Table 52 Traffic Congestion Impact on Route 89.2 (using scheduled running time data) 92 viii

9 ix LIST OF FIGURES Figure 1 the multiple regression model result Figure 2 Demonstration to Poisson passenger arrival with 30 stops for a route Figure 3 route 1 [18] Figure 4 the estimated impact cost of traffic congestion on passenger Figure 5 Travel congestion impact per Veh-hr Figure 6 Comparing Measuring travel congestion impact on buses due to lower average speed using AVL data & Scheduled running time data Figure 7 Route 23, MBTA, Boston [18]... 56

10 1 Chapter 1. Introduction 1.1. Overview Travel time of surface transit that shares right-of-way (ROW) with general traffic is affected by traffic congestion, leading to increased delay and unreliability. Since travel time and its reliability are the two most important determinants of travel mode choice, when transit suffers from traffic congestion, it makes transit less attractive, creating a vicious cycle with more cars on the road causing even more traffic congestion, making transit less and less attractive. Even though traffic congestion affects both public transit and private vehicles, the effect on transit is greater because transit vehicles can t change routes to avoid congestion. Thus, there is something unbalanced, even unfair, about private transportation imposing large delays on public transportation that share the same street. Ideally, cities should manage street vehicles so that transit is protected from congestion, offering the public a high quality service as an alternative to driving. Some cities such as Zurich and Brussels do this well [1]. American cities, for the most part, do not. An important first step in connecting this imbalance is developing a method to measure the harm that traffic congestion imposes on surface transit. As Daniel Moynihan said, We never do anything much about a problem until we learn to measure it [2]. Quantifying the costs that traffic congestion imposes on transit is important for measuring the benefits of implementing transit priority including physical measures (e.g., bus lanes) and signal priority. Also, knowing how much traffic negatively impacts transit can also be used to justify using fuel taxes from general traffic to subsidize transit.

11 2 At this moment, the transit industry is in the era of Big Data. Automated data collection systems for transit have almost become part of every large agency. It increases the flexibility, ease, and accuracy of analyzing operations. Automated vehicle location (AVL) systems and automated passenger counting (APC) systems have been used to measure transit performance in many respects. The goal of this research is to see how AVL and APC can also be used to develop a method to measure how well (or poorly) the road network serves transit by measuring the impact that traffic congestion has on transit operations and users Research Objective Transit systems routinely report performance measure such as total ridership, fraction of missed trips, fraction of trips that were on time, and number of safety incidents [3]. These measures report to the public how well transit is serving the customers. In the same manner, a measure is needed to report how well the traffic system serves public transportation. Thus, the objective is to develop a method that can be applied routinely as part of an annual report card to report how well the traffic system serves public transportation. The goal of this thesis is to develop a method for measuring incremental delay and unreliability due to traffic congestion and how its cost can be quantified using routinely available datasets, namely, schedule data, automatic vehicle location (AVL) data and automatic passenger count (APC) data. Quantifying the costs that traffic congestion imposes on transit can be used: a) As a tool to evaluate a city s accommodation for public transportation.

12 3 b) As a tool to present the need for and the expected benefits of transit priority, including both physical measures (e.g., bus lanes) and signal priority. c) As justification for spending more of the fuel tax on transit as a way of mitigating the harm that traffic does to transit. The general approach this study follows is to compare actual running time for periods with congestion against running time during a period with very little traffic such as the period between 10:00 P.M. and 6:30 A.M. With AVL data, we can see how much greater are mean running times and running variability during periods with more traffic congestion. The complication is to exclude the effect of the greater passenger demand that typically occurs during the more congested periods. For this purpose, methods were developed to estimate the effect that passenger demand has on both mean running time and running time variability, so that the incremental impacts of traffic congestion can be identified. 1.3.Thesis Organization This thesis is divided into seven chapters. Chapter 1 provides an introduction to the thesis objective and its approach. Chapter 2 reviews the related previous work on determining traffic congestion and its impact on the reliability of transit service with focus on using AVL data. Then, chapter 3 reviews all the data sources that were available for this research. Chapter 4 discusses the followed methodology of analyzing the AVL data in depth. Also, it discusses the found weaknesses of AVL data and finally it provides some suggestions to improve the automated vehicle location system at MBTA. Chapter 5 presents the research methodology and its application to 8 example routes. Chapter 6

13 4 presents the results of implementing the methodology on the sampled routes and a summary of the finding. Chapter 7 concludes the study and provides suggestions for future research.

14 5 Chapter 2. Literature review The topic of this thesis is to develop a methodology for estimating transit delay due to traffic congestion using routinely collected data (AVL). This chapter will first provide a review of prior research focused on measuring congestion, reliability and using AVL data. Section 2.1 presents the pervious proposed methodology of measuring traffic congestion. Section 2.2 discusses using AVL data to study a transit system. Section 2.3 discusses the reliability and its impact on users and operators. Finally, section 2.4 discusses the rationale for conducting this research by discussing what has already been examined and how the previous research sets up this research Measuring Congestion Quantifying transit delay due to traffic congestion has been explored by many researchers using different tactics to study travel time. The traditional method to study travel time is the floating car method. It involves collecting records while riding the bus or by from a car following a bus. The test vehicle technique is too costly to apply systematically because it is labor-intensive. Thus, some researches have worked to develop models for buses travel time as a fraction of cars General traffic travel time (Levinson, 1983 [4]; McKnight, Levinson, Ozbay, Kamga, & Paaswell, 2004 [5]). For example, McKnight et al conducted a study to determine the congestion impact on bus travel time by developing regression models that estimate bus travel time as a fraction of car travel time [5]. Then, the model was used to

15 6 estimate the proportion of bus travel time due to the increase in traffic time over freeflow conditions for car. Some Dutch transit agencies use the bus speedometer to measure traffic delay directly as part of an AVL system [6]. The Dutch system is programmed to write a record when the speed drops below a certain level (e.g. 5 km/h) and when it rises above that threshold, while excluding time while serving a stop [6]. To our knowledge, there is no U.S. AVL system that makes speedometer records in this way, making it impossible to measure delay directly AVL systems Automatic Vehicle Location (AVL) systems are computer-based vehicle tracking systems that rely on Global positioning system (GPS). In last years, AVL system became widely part of every large transit system around the world.the general concept of the AVL system is the same. However, the quality and level of the information that the AVL system can provide vary from an agency to other. Many researchers have explored the usage of AVL data for improving transit system; one of the fundamental papers is TCRP Report 113 (Furth et al. 2006) [7]. Furth et al. provides a comprehensive guidance for how AVL archived data system can be collected and used to improve the performance and management of the transit operations [7]. The MBTA s automated data collection system has been explored in some studies. For example, Cham (2006) used the MBTA AVL/APC data to develop a practical framework to understand bus reliability. He studied variability in the running time using time point data for Silver Line, Washington Street service. Even though the Silver Line is supposed to function as Bus Rapid Transit route, Cham found that variability of running

16 7 time is high. Other researchers used the real time AVL data for off-line analysis such as Gerstle (2009). He used AVL heartbeat data, which is free and open to the public from the MBTA, to explore buses position traces in real time. He used the real-time AVL published data to understand bus travel time variation which is the same as the archived heartbeat data, where the location of the bus is known every 60 seconds. However, there is a difficulty in determining the departure time for first stop and arrival time to last stop. Therefore, this kind of data can be used to explore the variability in running time for a route but the start and end time has to be estimated Reliability The reliability of transit service is critical to both operating agency and users. Abkowitz et al. defines service reliability as the invariability of service attribute which influence the decisions of the travelers and transportation providers. [10]. Reliability affects the travel mode choice and departure time for travelers [10]. In more details, reliability attributes of concern to transit users include waiting time, in-vehicle time, transfer time (missed connection) and seat availability [11]. Passenger waiting time is sensitive to the reliability of the service. Muller and Furth (2006) call the extra waiting time passengers suffer due to unreliability hidden waiting time [12]. They discussed how the short headway service is sensitive headway variability reliability, while long headway service is more related to high and low extremes of the schedule deviation distribution. Transit agencies incur greater costs when reliability is poor. To illustrate, with less reliability transit agencies tend to increase the allowed time in order to limit the probability that a trip will start late because of the lateness of the pervious trip [7].

17 Conclusion There are numbers of studies that have been conducted lately that are related to assessing the benefits of AVL data to study travel time or reliability of transit service. However, none of them specifically included measuring delay due to the traffic congestion and its impact at the route or system level.

18 9 Chapter 3. Data Sources Data used for this study was obtained from Massachusetts Bay Transportation Authority (MBTA). No special data collections were attempted. All the used data are the data that MBTA collects routinely. In this chapter, short description for the data sources that has been used in this research will be provided. The data sources are the automated vehicle location system and the automated passenger counting system Automated Vehicle Location system (AVL) Every MBTA bus is part of the MBTA automated vehicle location system, which is based on Global Positioning System. The AVL data is used in real time for operation control, incident management and delivering information to users. Ideally, automated vehicle location data is also archived for later analysis such as evaluating performance and revising running time schedules. The types of automated vehicle location data that were available to this research are heartbeat, time-point, and announcement records data. Each kind of AVL records will be described below Heartbeat Data Heartbeat data shows bus location every 60 second. The data also has the time when a location message stamp. The main use of heartbeat data is to provide the real time information for operators and users. Table 1 shows what information each record of heartbeat data includes.

19 10 Heartbeat is location-at-time data which, as Furth et al. (2006) show, is not well suited to archived data analysis. It will not be used to measure the running time for bus routes in this study for different reasons. First, departing first stop and arriving to last stop cannot be distinguished, and in the best-case scenario estimated running time could have two minutes error. Second, with heartbeat data the number of served stops per trip is hard to be obtained because it is difficult to distinguish between a stop that was for dwelling or for any other reason (such as traffic congestion). Table 1 Description of Heartbeat Data Data_Label (columns) Description logged_id Unique Number for each record (every 60 s) Calendar Date Message_type Null Latitude latitude of the bus location Longitude longitude of the bus location Adherence Adherence from schedule Odometer Odometer Validity Related to GPS signal Message_timestamp The Actual time for the record. Source_class N/A source_host Vehicle ID destination_class N/A destination_host N/A Time-point Data Time-point record is the most clear and straightforward type of AVL data. It is time at location data; time at location data shows when a bus passes a specific location (Furth et al. 2006). Time-points data shows arrival & departure time at key stops

20 11 (time-points) along a route, as determined by the bus computer using Global Positioning System. Also, it shows how long the bus was delayed from its schedule at each time-point. Table 2 lists and describes each column in a Time-point record. In order to inspect the quality of the data, we used the data to measure travel time for two MBTA routes (route 1 and 28). Unfortunately, the running time precision was high (unacceptable) at the first and last segment in both routes. My explanation for that is because of the complicated bus movements at terminals, stopping often at multiple locations, and also often going under a roof made the possibility of GPS error high. So, in some cases we got signal indicates that a bus departed the terminal but in actuality, the bus just was moving within the terminal. However, time-point data is an accurate tool to capture headway deviation and departure deviation at each time-point (except the first and last one) because in each row of the data there is scheduled and actual arrival time to a time-point. More details about what the time-points data contains are shown in the table below.

21 12 Column Name Table 2 Description of Time-point Data Description Crossing_ID Unique ID for each record of data Service_Date Service Date. District District (goes with run.) Run The four-digit run number Block The block of the piece of work. Operator Badge number of the driver. Vehicle vehicle identifier Half_Trip_Id Unique Number identifies that Half_Trip. Route Rout name Direction Direction ( Inbound or Outbound ) Variation Route Variation such as ( 39-3, 39-_, ) Stop Stop Number Time-point Time point abbreviation (5-letter code.) Time-point Order Order of time-point within HalfTrip (1 st time-point of HalfTrip would be 1, etc.) Scheduled Scheduled time Arrival Actual Arrival time to the Time-point Departure Actual departure time Earliness Deviation from departure time (s), in which positive is early and negative is late. Scheduled Headway Scheduled leading headway (seconds) Headway Actual leading headway (seconds) PointType Startpoint, Midpoint, Endpoint. StandardType Schedule, Headway, Express. Standard N/A Include N/A Announcement Record Data Announcement record data has every announcement made along a route with the location and the timestamp. The records show whether the announcement was made internally or externally and whether it was made visually or audibly and also whether the door was open or not. However, there is no clear identification of trips on the data and no matching to schedule. More details about what the announcement record data contains is shown in the Table 3.

22 13 Column Name ROUTE_ABBR ROUTE_DIRECTION_NAME ANNOUNCE_DESC LOGGED_MESSAGE_BS1_ID CALENDAR_ID MESSAGE_TYPE_ID MESSAGE_TIMESTAMP LOCAL_TIMESTAMP SOURCE_HOST LATITUDE LONGITUDE ADHERENCE ODOMETER VALIDITY MDT_BLOCK_ID MDT_RUN_ID EFFECTVE_SERVICE STOP_OFFSET CURRENT_DRIVER ROUTE_OFFSET DIRECTION LONG_FIELD_1 LONG_FIELD_2 LONG_FIELD_3 LONG_FIELD_4 LONG_FIELD_5 Table 3 the Components of Announcement record Data Column Description Route abbreviation Inbound or outbound announcement description (e.g a name of a terminal) Unique identifier for each record. Date Null Actual time when the record made (GMT time) Actual time when the record made (local time) Vehicle ID LATITUDE LONGITUDE Schedule Adherence Odometer Associated With GPS signal Associate with the scheduled block Associate with the scheduled block N/A It is a bus s relative location in an ordered list of all stops for the given run or block. Driver ID Related to the given MDT_RUN Direction ( Inbound or Outbound ) Identifier for each announcement. N/A Identifier to how the announcement was played. [e.g 545 indicates the announcement gets played externally (this occurs when the front doors are first opened and every 30 seconds after that if they remain open; the announcement gives the route and destination); 257 indicates the announcement gets played internally both audibly and visually (this occurs as the bus approaches a stop); 273 indicates the announcement gets played internally visually only (this occurs when the front doors are first opened and every 30 seconds after that if they remain open)]. Null Null BYTE_FIELD_1 to 5 Null Announcement records data will be used in this research in order to obtain travel time between the first stop and the last stop. It is possible to capture last announcement made at the first terminal, when the bus door was open, by tracking the records in the

23 14 LONG_FIELD_3 column. Also, announcement records data can be used to provide the number of stops that were serviced per trip, as it will be described in depth in the next chapter Automated Passenger Counting Data (APC) Automated passenger counting data that will be used in this research are the MBTA APC summary report and ride check data for individual trips. The fraction of buses that have APC is around 10% of the buses. The APC summary report contains the average of passengers boarding and alighting for every scheduled trip, obtained from many observations. Ride check data has the number of passengers getting on and off at each stop along with opening and closing door times for individual trips. The sample size for ride check data is not large compared to APC summary report. Ride check data is used to calculate the percentage of passengers boarding at first stop and alighting at last stop, since those passengers movements don t affect the running time as measured from the announcement data (It will be described in depth in chapter 4). Ride check data will also be used to estimate a stop time model, as described in chapter 5.

24 15 Chapter 4. AVL Data Analysis Methodology As described in chapter 3, the AVL data that were used are announcement records and time-point data. In this chapter, the methodology of analyzing these data will be described Announcement record data processing Announcement records will be used to determine the travel time for a bus between departing first stop and arriving to last stop. Also the data will be used to determine the number of stops that were served during each trip. The size of the data is large, for example, more than 2 million records were archived for Route 1 in three months only. The announcement record data also includes records that are not useful, including stop requested announcements, out of service announcements and some repeated records. For that, we wrote a code using Python to identify complete trips and determine each trip s running time and the number of stops made during each trip. The announcement data doesn t include a trip ID. So, there is no field that gathers the data by trip. However, each announcement record has a vehicle_id, a driver_id, a unique ID and the direction of the bus whether it was running inbound or outbound. Thus, the code s main mechanism is to track each vehicle with its direction. Starting when the bus was at first stop and ending when it reached last stop, which helps distinguishing each trip records. Basically, the code tracks each bus to get the last announcement played in the bus when the door was open at the first stop; that moment is considered as the departure time at the first stop. Then, we track the bus until we find the first announcement made when bus s door was open at last stop, and consider it as the arrival time to the last stop.

25 16 The maximum possible time error for this methodology is 30 seconds because the announcements we tracked are played when the front door is first open and every 30 seconds after that if they remain open. A bus with an open door was considered to be at a stop if it was with 50 m of the standard stop location. The results were viewed using ArcGIS in order to confirm that the stop locations were within the acceptable range. Some criteria were added to the logic to confirm a complete trip and exclude odd data records. 1. If a trip announcement records are interrupted with out of service announcement, the trip was eliminated. 2. If the driver s ID changed in the middle of a trip, the trip was eliminated Time-point data processing Time-point data is used to find the effect of the running time variability on passengers as it will be described in the methodology chapter. Time-point data analysis process (code mechanism) o For each individual trip, assign departure time from first stop as the trip time. o Classify the trip date into either (weekday, Saturday or Sunday schedule) [Taking in consideration holidays schedules]. o Calculate ideal and observed headways at each time-point. o Grouped trips according to their trip time. o Find arrival time deviation from scheduled time for each trip at each timepoint. (Also, departure deviation )

26 17 o Classified each period to high /low frequency service according to its average actual headway at the second time-point. The analysis code was written using Python by using data analysis libraries (Pandas and Numpy). We ran the code to analyze four months data records; the data cover all the trips for MBTA from Aug 31 to Dec 27/2013 for all the chosen routes Evaluation & Suggestion for the Reviewed AVL Archived Data The information that was obtained in this research from the AVL data is very useful but not all what we ideally would expect. The system might have been designed in a way that all that an analyst would need for off line analysis is time-points records. This section will suggest some changes in order to make the AVL system data more useful for off line analysis. The found weaknesses and the proposed improvements are The announcement record data can be used to show where and when the door is open but the name of the stop is not given unless the stop was terminal or timepoint. Thus, creating unique announcements for each single stop is suggested, or at least noting the stop ID as a part of the announcement record. We have noticed in the data that some drivers turn the external announcement to out of service before reaching the last terminal in order to show passengers who are waiting at the last terminal that they can t get into the bus at that time (because there is shift change or any other reason). This is a problem because it is hard to distinguish between if the out of service announcement was used to show passengers that they couldn t get into the bus or if the bus was really taken out of service. Thus, creating new external announcement that the driver can use in

27 18 specific situations other than out of service like no boarding allows at this time is suggested. It is hard to match announcement record data with time-point and heartbeat data for a trip. Thus, creating a half_trip ID for each trip in all the datasets (Announcement record and heartbeat data) is suggested. Although leaving first stop and arriving to the terminal are critical events, current time-point data doesn t capture the well. Thus, we suggest creating a unique announcement for first terminal departure (its coordination should be set on a point where the bus cannot stop and where leaving the stop is guaranteed) and also creating an announcement for the arrival to last stop. During the research, I attempted to analyze Route 111. There was a difficulty analyzing the trips that run (inbound) because there are two variations with different starting points for the route and they have the same external announcement even though the trips don t start from the same places. Thus, the announcement record needs a field to distinguish each route variation (time-point data already has this feature).

28 19 Chapter 5. Methodology The primary analysis will be based on the AVL data (plus some APC data). This chapter describes the stop time model in section 5.1 and the data organizing/grouping methodology in section 5.2. Then, traffic congestion impact will be discussed at two different levels; mean running time and variability in running time impact in sections 5.3 &5.4. Value of time for passengers and operating cost is discussed in section 5.5. Then, section 5.7 measures traffic congestion using a simple method that uses schedule data instead of AVL data. Finally, the methodology will be applied and demonstrated in detail for one route, Route 1 in section Stop Time Model Stop time consists of acceleration delay, deceleration delay, dwell time, opening/closing door time, and return to traffic time. Dwell time, the longest component, is most directly related to demand. However, the other components are hard to measure using available automated data. Reasonable assumptions will be made where we try to model those components at a low traffic period. However, from AVL data a dwell time model was estimated Dwell Time Model Since ride check data has the opening and closing door time as well as number of passengers getting ons and offs by stop, a multiple regression model can be performed to estimate the dwell time, defined as the time between doors open and doors close, as a function of the numbers of boarding and alighting passengers.

29 20 Figure (1) shows the multiple regression model results. Estimates are measured in seconds. 14,289 observations were used in the regression. The coefficient of determination or R 2 is 0.64, which means the model explains 64% of the variability.the model shows as 9.79 s standard error against a mean dwell time of s. Looking to the t-test and p-value, all variables are statistically significant. The F-value shows that the model overall is statically significant. SUMMARY OUTPUT Regression Statistics Multiple R R Square Adjusted R Square Standard Error Observations ANOVA df SS MS F Significance F Regression Residual Total Coefficients Standard Error t Stat P-value Lower 95% Upper 95% Intercept ON OFF Figure 1 the multiple regression model result Using these results, dwell time will be estimated using equation (5.1) Dwell Time = Tons * Toffs *1.87 (5.1)

30 21 At, the trip level, the part of dwell time dependent on passenger activity can be called passenger service time (Pax ST ),given by Where Pax ST = 4.35 Tons Toffs (5.2) Tons = total passengers who boarded at trip, excluding the first stop. Toffs = total passengers who alighted at trip, excluding the last stop. Passenger activity at the first and last stop is excluded because as running time is measured from announcement data; dwell time at the first and last stop is not Lost time Increased passengers activity most directly affects running time by increasing passengers service time. In addition, it leads to buses stopping more often. Therefore, it is necessary to estimate the parts of stopping time that are independent of passenger activity which we call lost time. Lost time has several components, listed in Table 4. - Lost time when door open, from regression model, is 1.6 s. It s extra time for the first passengers to board or alight, and for the drives to close the door after the last passenger. - Other lost time is assumed to be 2 second. That includes time between wheels stooping and door opening, between door closing and wheels rolling, and waiting to return to traffic. - Acceleration and deceleration delay. Because this study focuses on finding the effect of traffic congestion, time for a bus to accelerate, decelerate and return to traffic at the low traffic period will be estimated to equal to the needed time

31 22 Where and any extra time that might occur at the other periods (e.g. peak periods) will be considered as delay due to traffic congestion. The model that will be followed on this study, assumes uniform acceleration (equation 5.3). D acc = t acc t unacc = V a acc V = V (5.3) 2 a acc 2 a acc D acc = the acceleration delay t acc = time needed to accelerate t unacc = time needed to pass over the acceleration distance without stopping V = speed of non-stopping vehicle = the average acceleration rate a acc We assume that the applied acceleration rate is 1.5 mph/s and the nonstopping bus runs at speed of 21 mph. Then the acceleration delay is 7 seconds. The speed of 21 mph is mix of the buses that would otherwise pass at full speed (30 mph) versus at a low speed due to signals and queuing. The deceleration rate is assumed to be twice the acceleration rate. Thus, the deceleration delay equals to half of the acceleration delay (3.5 second). o Together, then, lost time per stop is assumed to be 14.1 second (see Table 4). Table 4 lost time components and their values. Description Value Deceleration Delay 3.50 Lost Time With door open" 1.60 Acceleration Delay 7 Other 2 Total Lost time 14.10

32 23 For trip as a whole, total lost time is Where Lost_T Trip= 14.1 *( N_stop -1) (5.3) Lost_T Trip = lost time due to making stops for a trip. N_stop = number of served stop at a trip. In our case study, I used (N_STOP -1) because captured running time for each trip starts from closing the door at first stop and ending when the door is open at last stop Grouping trips Trips will be grouped into different periods of the week with relatively homogeneous running traffic conditions and demand. For the case of MBTA, we developed the set of periods show in Table 5. Table 5 data grouping periods. Interval Day & Period Weekday Saturday Sunday Interval 10:00 PM 6:29 AM 0Low_Traffic 6:30 AM 6:59 AM 5Shoulder/Eve. 10:00 PM 7:59 AM 7:00 AM 8:59 AM 1AM Peak 6Sat_Morning 7SUN_Morning 8:00 AM 11:59 AM 9:00 AM 1:29 PM 2Midday Base 1:30 PM 3:59 PM 3Midday School 8Afternoon_WKEND 12:00 PM- 5:59 PM 4:00 PM 6:29 PM 4PM Peak 6:30 AM 9:59 PM 5Shoulder/Eve. 9Evening _WKEND 6:00 PM 9:59 PM

33 Average lower speed impact Using announcement record data, running time (RT) p and the number of stops served (N_stops) for each trip are obtained as discussed in section (4.1). Then, mean adjusted running time for a period after accounting for demand, which means excluding serving passengers and serving stop effect out of observed running time, is Adj(RT) p = RT p 14.1 ( N_stop -1) Tons p Toffs p (5.5) Where ; Adj(RT) p = Adjusted mean running time for period p. RT p = mean running time for period p. N_stop = average number of served stop at a trip. Tons p = Average total passengers who boarded at trip, excluding the first stop (for period p). Toffs p =Average total passengers who alighted at trip, excluding the last stop (for period p). Thus, if running time of buses that run in mixed traffic is combination of a) Needed travel time b) Stop Time c) Delay due to traffic congestion d) Inherent randomness e) Operational control Then, RT p includes all five components a,b,c,d and e. Adj(RT) p eliminate component b. And Adj(RT) o (for low traffic period 0) also eliminates component c. Thus, the estimated incremental mean running time Del(RT) p for a period p due to traffic congestion is the difference between its Adj(RT) p and adj(rt) o for the low traffic period (see equation 5.6).

34 25 The reason why we use the adjusted running time at low traffic period as a base is that it gives the needed travel time during non-congestion time plus an estimation of all the inherent factors that affect net travel time for a bus such as road condition, drivers behavior and other factors that affect net travel time. Del(RT) p = Adj(RT) p Adj(RT) o (5.6) Negative value means bus consumed more time at the low traffic period than the other periods, which indicates poor operating management for a route. Del(RT)p reflect impact to operating agency.on the other hand, the impact on a passenger depends on the extra running time for the segment of route that his/her trip covers. Ideally, load per segment can be obtained using on-off data. Then, weighted segment level running time can be calculated in order to get total pax-min. Thus, the ideal method demands data collection are:- A. Running time by stop (treat each stop as time-point, plus record door opening / closing) B. Load by stop (requires balancing ons & offs) Unfortunately, the ideal method can t be implemented due to lack of data. However, the method that will be followed in this research is to assume that passenger trip covers a certain fraction of the route. Then, the estimated increase on a passenger mean travel time due to traffic congestion (Del(TT) p ) will be equaled to the percentage of route length a passenger trip cover (% RT_L ) multiplied by the estimated incremental mean running time ( Del(RT) p )at period p:- Del(TT) p = % RT_L * Del(RT) p (5.7) For our case study, following Furth (1998), a passenger trip on average covers 40 % of the route length [16].

35 Variability in Running Time impact Travel time variability leads to service unreliability, which imposes costs on both transit users and transit agencies. In one hand, variability in the travel time affects operating cost because it leads to an increase in trip s allowed time *. On other hand, service unreliability increases passengers waiting time and their potential travel time [12]. Variability in running time for the entire trip uses to determine the variability impact on operating cost. However, passenger trip cost is more associated with variability at stop level. Thus, for our case study, announcement record data will be used to study the variability at the trip level & time-point data will be used to study the variability at the stop level only for the reasons that were discussed in chapter 2 and Variability at the trip level By Variability at the trip level, we refer to the variability in the running time between first and last stop Transit service that runs in mixed traffic has variability in the running time, which is unavoidable but controllable. For that, transit agencies uses different scheduled running time for different times of the day in order to reduce the effect of variability in running time on passengers and operating cost. Thus, the observed running time at period p, could contain scheduled and unscheduled running time variability. The following section ( ) will discuss how the unscheduled variability can be captured. * Allowed time is scheduled running time plus recovery time for a trip.

36 Variations from the scheduled running time V FSch (RT) The study period might include more than one scheduled running time (meaning we have variation in the scheduled running time at that period). Thus, in order to identify the traffic congestion impact on reliability, we have to identify the unscheduled variation at each period. In other words, the impact of traffic congestion on service reliability can be known by knowing how much the actual travel time varies from the scheduled using equation (5.12) Let, Where The observed travel time are (t ipj ) The scheduled travel times are (S ip ). i = scheduled trip. j = day. p = period of day. t = observed travel time. s = scheduled travel time. t p = the mean observed running time for a period. S P = the mean scheduled running time for a period. V Fsch (RT) P = estimated variance of the running time from scheduled running time in period p. V(S) P = variance of scheduled travel time in period p. V(RT) p = variance of (observed) travel time in period p. Then, the observed variation in the running time for a period is V(RT) P = 1 j i n 1 (t 2 ipj t p ) (5.8)

37 28 The variation in the scheduled running time for a period is V(S) p = 1 j n 1 i (S ip S p) 2 (5.9) Ideally, we would estimate the mean squared deviation from scheduled running time, which we call the variance from scheduled running time, for a given period p, by V Fsch (p) = 1 i (t 2 n 1 i s i ) (5.10) Where the sum is over all trips in period p. However, because AVL data used to measure running time (announcement data) doesn t indicate trip number, it is not possible to match each trip with its pair trip that happened next day and the following day and so on. By using unmatched data, the variation from scheduled running time for a given period can be calculated as follow V FSch (t) = E(t i s i ) 2 = E{[(t i μ T ) (s i μ S ) + (μ T μ S )] 2 } = E(t i μ T ) 2 + E(s i μ S ) 2 2E[(t i μ T )(s i μ S )] (5.11) (Expanding the square yields two other cross-product terms, but their values are zero.) Rewriting the last term in terms of r ts, the correlation between scheduled and actual running time in the period of analysis, and replacing population mean and variance with sample mean and variance where available, variation from scheduled running time is given by V FSch (RT) P = V(RT) P + V(S) P + (S P t p) 2 2r ts V(t) V(s) (5.12)

38 29 Because this correlation cannot be measured with unmatched data, it must be assumed. It is both conservative (i.e., to avoid overestimation) and reasonable to assume strong correlation, since the aim of the scheduling function is to schedule longer running times when actual running times are longer. For the following case study, the assumption of r ts = 0.8 will be followed Adjust running time variation for greater demand Greater demand affects running time by affecting the number of passengers and number of stops. In order to adjust running time variation for greater demand, their variability should be determined and its effect subtracted from the running time variability. Let, Serve Passengers Effect [V serve (RT) p] is the part of the running time variability that due to the variability in the number of boarding passengers per trip if we assumed passenger arrival process followed Poisson distribution. then, V serve (RT) p = (OnTime + OffTime) 2 * (boarding per trip) (5.13) Where OnTime = the estimated time needed for a passenger to board = 4.35 S. OffTime = the estimated time needed for passenger to alight =1.85 S.

39 30 Stopping Effect, V stops (RT) p, is the portion of the running time variability due to variability in the number of stops made per trips. V stops (RT) p = V(n_Stop) p * (LostTime) 2 (5.14) Where V(n_stop) = variation on the number of served stop. Lost Time = lost time per stop = 14.1 s. Note: V(n_stop) directly measured from AVL announcement data. Thus, equation (5.15) shows the variability in running time that is not due to the effect of greater demand; the need to make additional stops, and longer service times. AdjV FSch (RT) p = V FSch (RT) p V stops (RT) p V serve (RT) p (5.15) Impact on Operating Cost Running time variability affects operating cost because it is associated with layover time. The allowed time for cycle or half-cycle for a route is equal to the running time plus the layover time (or the recovery time). Furth et al. (2006) mentions that in order to limit the probability that a bus finishes one trip so late that its next trip starts late, allowed time for a route should be based on an extreme value such as the 95th-percentile running time [7]. While some of the agencies apply the 95 percentile rule at the round trip level, in this research we apply it to a one direction trip level (half-cycle) in order to match the MBTA methodology (for setting allowed time).

40 31 So, to find the extra deviation that due to traffic congestion: Let s consider that in any period, observed running time variation from the schedule for a period is a combination of: a) Variation due to inherent randomness (e.g. arriving signals just after they turn red; wheelchair use). b) Variation due to operational control (e.g. whether dispatch is on time; holding at time-points; difference in speed, acceleration, braking between operators). c) Variation due to serve more or fewer passengers. d) Variation due to making more or fewer stop. e) Variation due to traffic congestion. Then, o AdjV FSch (RT) p for period p represents components a,b and e. For period 0, AdjV FSch (RT) o represents components a and b. Then, Incremental variation due to traffic congestion (V e,p ) is equal to the difference between adjusted running time at period p and at low traffic period. V e,p = AdjV FSch (RT) p - AdjV FSch (RT) o (5.16) Then, the estimated incremental recovery time (DelRec) can be represented as equal 1.64* sqrt[variation due to traffic congestion] as it is presented in equation (5.18). Recovery p =1.64 V FSch (RT) p (5.17) DelRec p = 1.64 { V FSch (RT) p V FSch (RT) p V e,p } (5.18) It is about 95 percentile (with assuming random distribution)

41 Variability at stop level The purpose of studying the variability at the stop level is to obtain the effect on passengers. Passenger s reactions toward variability in running time is to budget more than the expected trip time in order to achieve a certain level of confidence to arrive to his/her destination not late. Thus, variability in running time of buses affects passengers riding and waiting time. Since passengers arrival process to a stop is independent on published scheduled with high frequency service but dependent on schedule with low frequency service, the impact of variability on the waiting differ in each case as it will be described later in section ( & ). However, passenger s behavior towards variability on travel time is the same with low or high frequency service as it will described in section ( ). Our definition of high frequency service is that a bus route is operated with average headway equal to 13 min or less at that period Impact on waiting time with high frequency service For high frequency service, the focus is on headway deviation because passenger arrival process is independent of schedule departure time and passengers always aim to the next trip. It is well known for such a case that the expected waiting time for passengers is given by E[W] = h V(h) (1 + 2 h 2 ) (5.19)

42 33 The part that stems from variability is excess waiting time, given by h 2 (V(h) h 2 ). V(h) is partly due to traffic but also partly due to demand. Assuming on-time dispatching and independence in running time between two successive trips a and b, where H b = t b -t a and so,v(h) = 2V(t) t = running time from terminal to boarding stop involved variability generated from demand; variability in stopping time and service time over this segment. Because waiting time is dependent on the boarding stop, O-D matrix is needed to capture this level of data, which is not available for this research. In order to solve for this problem, we assumed that boarding stop is at a point at which 25% of route-level stopping and 33% of route-level service time has occurred (33%, not 25%, because boarding take more time than alighting and tend to be concentrated near the start of the route); then the adjusted headway variance for a given period is AdjV(H) p = V(H) p 0.5 [V stops (RT) p V stops (RT) o ] 0.66 [V pax (RT) p V pax (RT) o ] (5.20) Since low-traffic periods have no short headway service that can be used as a base for determining headway variation under low traffic conditions. However, the existence of headway variation on MBTA rapid transit lines indicates that there are causes for headway variation apart from traffic congestion, such as failing to start on time and variability in the speed operators choose. A reasonable assumption for headway variation

43 34 in the absence of traffic is 1 min 2. Therefore, the incremental excess waiting time for a period is DelWait p = H 2 AdjV(H) p 1 H 2 (5.21) Impact on waiting time with low frequency Service On other hand, for low frequency services, the focus is on departure time deviation from the schedule, because passengers arrival is dependent on the published schedule. Following Furth & Muller (2006) [12], we assume that passengers limit their chance of missing the bus by arriving no later than the 2-percentile departure time (relative to the schedule that is the 2-percentile departure deviation). Thus, passengers excess waiting time is then the difference between 2-p departure deviation (DepDev 0.02) and mean departure deviation (DepDev mean ) from the boarding stop, as ExcessWait = DepDev mean DepDev 0.02 (5.22) Departure deviations at time-points can be calculated from time-point data. Since boarding tend to occur in the earlier part of a route, choose time-point closest to 25% of the way along the route. Because excess waiting time applies only to lower demand periods that do not require high frequencies, no adjustment is made for demand. The impact of traffic on excess waiting time can be estimated as

44 35 DelWait p = ExcessWait p ExcessWait o (5.23) where ExcessWait p = excess waiting time at period p ExcessWait o = excess waiting time at low traffic period Impact on potential (Budgeted) Travel Time Potential travel time is the extra time that people budget for travel but on average do not consume it. In order to limit the probability of arriving late at a desired destination, we assume that people budget for arriving at the 95-percentile arrival time at destination stop. Potential travel time is therefore Arr 95pct Arr. This can readily be calculated from time-point data; however, it is not amenable to adjustment to account for demand, and therefore it is difficult to isolate what part of it can be attributed to traffic congestion. An approximate way to estimate potential travel time is to take the difference between the 95-percentile arrival time at a stop and the scheduled arrival time. If that step were the terminal this would be exactly the same as the needed recovery time. On average, a passenger s destination stop will be toward the end of a route but not at the very end; we will assume that Potential travel time (Potential p ) is 75% of recovery layover, as Potential p = 0.75* Recovery p = 0.75 * 1.64 V FSch (RT) p (5.24) Therefore the incremental potential travel time due to traffic congestion (DelPotential p ) is DelPotential p = 0.75* DelRec p (5.25)

45 Value of time According to the NTD website (data of 2012), the operations cost per bus/hr is $/hr for MBTA [13]. However, traffic delays leave vehicle-miles unchanged. So, in this study 70% of the bus operation cost ( $/hr) will be used to evaluate the cost of the delay due to traffic congestion since traffic delays leave vehicle miles unchanged [17]. U.S. DOT recommends using 12 $/h travel time values for personal and local trip (within the city)[14]. The value is intended to reflect average cost of travel time value which is between 35-60% of average wages. Potential travel time is the extra time that people budget for travel but on average do not consume it. Thus, it is reasonable to give it a unit cost less than the unit cost of invehicle time. (Otherwise, people arriving at a terminal before their budgeted time would sit in the bus until their budgeted arrival time.) Following Furth & Muller (2006), a unit cost of 75% of the value of in-vehicle time is assumed. The ratio of waiting time cost to in-vehicle time should be more than 1. Therefore, the cost of excess waiting time will be considered 150% of the cost of in-vehicle [12]. Finally, to get the annual impact, we calculated for 56 Saturdays, 56 Sundays and 253 weekdays per year, based on the schedules the agency operates on holidays.

46 Incremental cost due to increased (Variability) Incremental cost due to increased (Average delay) Summary Traffic congestion increases cost to transit operators by increasing mean running time and running time variability. It increases costs to transit passengers by increasing riding time and by increasing service unreliability. This leads to increases in waiting time and potential (budgeted) travel time. The components of cost impact for a given period, are summarized in Table 6. The overall cost is then obtained by aggregating over periods. Table 6 Annual cost component summery for a period p Impact on operating agency Impact on passengers DelRT p *trips p * $ DelRT p * (Ons/ trip p ) * trip p *$12 DelRec p *trips p * $ (DelWait p * DelRec p *0.75)* (Ons/ trip p ) * trip p *$12 Where trips p = numbers of trips per year in period p. ons/trip p = boarding per trip in period p. DelRT p = incremental running time, period p. DelRec p = incremental recovery time, period p. DelWait p = incremental waiting time, period p.

47 AVL-Free Methodology In this section, we also test to what extent the traffic congestion and its associated cost can be estimated using schedule data instead of AVL. The rationale for using the schedule running time data is based on the fact that scheduled running time ideally should reflect the observed running time (from AVL). Travel time will be obtained from the scheduled running time. The number of stops made by a bus during a trip will be modeled, as it will be described in the following section. However, running time variability and its associated impact cannot be captured from schedule data only The Number of Stops Made Model. The number of stop can be known by using AVL data [what we did]. However, without AVL data, we can model the number of stops made based on a Poisson passenger arrival. Let, m = movements (ons/offs) per stop m = 2 Ons/trip N (5.26) Where N= number of stops. Then, the probability of a bus stopping at a given stop equals the probability of one or more people waiting to get on or off: P(stop) = 1 e m (5.27) Thus, the expected number of stops made in a trip is Stops Made = N (1 e m ) (5.28)

48 Stops Made 39 Figure 2 Demonstration to Poisson passenger arrival with 30 stops for a route Summary Ons/trip Using scheduled running time will require following the same process as that we used for AVL data, except that the scheduled running time is used instead of observed running time (AVL); and the number of stops made is modeled instead of being observed. The only impact that could be captured is the one that is due to lower average speed. The accuracy of the traffic congestion impact using secluded running time depends on the quality of the operations management for the transit service.

49 Application to MBTA Route 1 Route 1 is one of the key bus routes in Greater Boston area. It runs between Harvard Square and Dudley station. Harvard to Dudley is considered the inbound direction (see figure 3). Figure 3 route 1 [18] The first step of analyzing Route 1 is to obtain the number of planned trips per day from the published schedule in order to weight the results of a period at different days (see Table 7).

50 Direction Period Day Scheduled Trip/day Scheduled Trip/year Scheduled Trips/year Share of year s trips/period Share of year s trips 41 Table 7 trips throughout a year for Route 1, MBTA SAT % IB 0Low_Traffic SUN % WKDY % Total % 20% SAT % IB 8Afternoon_WKEND SUN % Total % 9% SAT % IB 9Evening _WKEND SUN % Total % 5% IB 1AM Peak WKDY % 9% IB 2Midday Base WKDY % 14% IB 3Midday School WKDY % 9% IB 4PM Peak WKDY % 13% IB 5Shoulder/Eve. WKDY % 17% IB 6Sat_Morning SAT % 3% IB 7SUN_Morning SUN % 2% OB OB OB 0Low_Traffic 8Afternoon_WKEND 9Evening _WKEND Total % SAT % SUN % WKDY % Total % 22% SAT % SUN % Total % 8% SAT % SUN % Total % 5% OB 1AM Peak WKDY % 8% OB 2Midday Base WKDY % 13% OB 3Midday School WKDY % 9% OB 4PM Peak WKDY % 13% OB 5Shoulder/Eve. WKDY % 17% OB 6Sat_Morning SAT % 3% OB 7SUN_Morning SUN % 2% Total %

51 Then, the results are weighted based on the number of trips per week (See table 7). Then, we processed the data we have (different data sources) in order to get the needed inputs for the proposed methodology (See table 8). Table 9 shows the first part of the methodology, where we adjusted all the needed values to capture traffic congestion impact. Then, table 10 shows the annual impact of traffic congestion in $ and hours. The used AVL data for the route covers the entire fall 2013 period trips (August 31 December 27). Note, abbreviation are defined in Appendix Table 8 the data sources and its outputs (Data_Source) Output APC Data Pax/trip %ons at First Stop % offs at last stop Announcement Record data Running time (RT) S(RT) standard deviation of (RT) (nstop) number of the serve stop during a trip S(nStop) standard deviation of (nstop) Scheduled Running time Running Time (RT ) Standard DEVIATION (RT ) Published Schedule Trip/yr Time-point_Data Headway Class (short, long) Headway mean Headway Standard deviation. Expected Departure Deviation 0.02 percentile departure deviation

52 Rt. Dir. Period hdwy netons netoffs RT nstop S(nStop) AdjRT AdjTT SFsch(RT) AdjSFsch(RT) (H) S(H) AdjS(H) DepDev DepDev Wait AdjPot'l 43 Table 9 Route 1 variables and corresponding adjustments (using AVL data) 1 IB 0 LH n/a IB 1 SH IB 2 LH n/a IB 3 SH IB 4 SH IB 5 SH IB 6 SH IB 7 LH n/a IB 8 SH IB 9 LH n/a OB 0 LH n/a OB 1 SH OB 2 LH n/a OB 3 SH OB 4 SH OB 5 SH OB 6 SH OB 7 LH n/a OB 8 SH OB 9 LH n/a Negative sign indicate early departure

53 Annual Impact Note, for more details about the unit of cost see section 5.5. The following equation was used to calculate the cost of the impact of traffic congestion on passenger and operating. Cost ($/yr) = Unit cost *[N_E/Y] * [ TCI] (5.29) Where Cost = cost of traffic congestion impact per year Unit cost = ($/hr) N_E/Y = number of effected trips or passengers TCI= Annual traffic congestion impact (h/yr) In the case, when the traffic congestion impact values were negative, values were set to be equal zero. The total incurred loss due to traffic congestion on passengers and the agency is shown in table (10). The results show that the estimated annual traffic congestion cost impact on operating cost is $1.2 million and the cost impact on passengers is $6 million. The cost of unreliability due to traffic congestion on both passengers and operating cost is around $ 4.1 million/year.

54 Dir period hdwy trips/yr Ave (ons/trip) DelRT (min) DelRecov (min) DelTT (min) DelWait (min) DelPot'l (min) Total 45 Table 10 Traffic Congestion Impact on Route 1, MBTA, Boston. IB 0Low_Traffic Long_Headway IB 1AM Peak Short_Headway IB 2Midday Base Long_Headway IB 3Midday School Short_Headway IB 4PM Peak Short_Headway IB 5Shoulder/Eve. Short_Headway IB 6Sat_Morning Short_Headway IB 7SUN_Morning Long_Headway IB 8Afternoon_WKEND Short_Headway IB 9Evening _WKEND Long_Headway OB 0Low_Traffic Long_Headway OB 1AM Peak Short_Headway OB 2Midday Base Long_Headway OB 3Midday School Short_Headway OB 4PM Peak Short_Headway OB 5Shoulder/Eve. Short_Headway OB 6Sat_Morning Short_Headway OB 7SUN_Morning Long_Headway OB 8Afternoon_WKEND Short_Headway OB 9Evening _WKEND Long_Headway Total h/yr, INBOUND 4, , , , ,072.2 Total h/yr, OUTBOUND 3, , , , ,950.3 Total h/yr 7, , , , ,022.5 unit cost ($/h) Impact ($/yr) 800, ,699 2,297,076 2,045,403 1,692,202 7,249,376

55 Rt. Dir. Period RT Stdv(RT) Ave_Pax N_STOP Stop_time Net T.T. Delay /trip TT.Dealy Analyzing route 1 using scheduled RT. In this section, the scheduled running time is used instead of AVL data. Table 11 Route 1 variables and corresponding adjustments (using scheduled RT) min min PAX min min min min 1 IB IB IB IB IB IB IB IB IB IB OB OB OB OB OB OB OB OB OB OB

56 Dir. Period Delay /trip Trip /yr Operation Cost Pax_delay (min) Esti. Pax Delay cost 47 Table 12 Traffic Congestion Impact on Route 1(using scheduled running time data) min/trip $/yr min/trip Pax/yr $/yr IB 0-7, ,861 - IB ,289 40, , ,718 IB ,060 80, , ,829 IB ,289 84, , ,409 IB , , , ,151 IB ,072 62, , ,032 IB ,176 10, ,485 22,691 IB , ,146 7,300 IB ,192 53, , ,173 IB ,736 20, ,798 40,140 OB 0-8, ,318 - OB ,036 41, , ,928 OB ,060 71, , ,081 OB ,542 72, , ,369 OB , , , ,408 OB ,325 39, ,626 89,440 OB ,232 10, ,414 18,458 OB , ,702 4,462 OB ,192 44, , ,390 OB ,736 14, ,004 30,817 IB Inbound-Total 497,977 1,415,444 OB The route _Total 401,176 1,160,353 Total 74, ,153 4,055,344 2,575,797 The results show that due to lower average speed, the estimated annual traffic congestion cost has impact on passengers about 2.3 million and 800 thousand on operating cost. Using AVL data, the corresponding figures were 2.5 million and 900 thousand, showing that, at least for this route, using schedule data yields a very similar estimate. Of course, using schedule data in place of AVL doesn t allow one to estimate impacts due reliability.

57 48 Chapter 6. Results In this research we attempted to analyze 9 bus routes at Greater Boston area that have different demand level (see table 13). The used AVL data for the routes covers the entire fall 2013 period trips (August 31 December 27). Table 13 list of the chosen routes Route Def. Destinations (two ends stops) 1 Key route Harvard Square Dudley station 23 Key route Ashmont Station - Ruggles Station 28 Key route Mattapan Station - Ruggles Station 39 Key route Forest Hills - Back Bay. 66 Key route Harvard square - Dudley Station 77 Key route Arlington Heights - Harvard Station 9 Regular route Copley square and City Point Bus Terminal 89 Regular route Clarendon Hill - Sullivan Station 89.2 Regular route Davis Square - Sullivan Station 99 Regular route Boston Regional Medical Center -Wellington Station The annual impacts of traffic congestion on each of the routes are shown in the table below (table 14). The impact on the key bus route tends to be similar. The estimated average impact cost on a key bus route is equal to 6.8 million /year. The impact on operating cost is 1.14 million/year on average. The impact cost of unreliability alone is around 280,000 on operating cost and 3.3 million on passengers (see Table 14).

58 49 Table 14 Summary Annual Impact of Traffic Congestion on the chosen routes Impact due to lower Variability impact Total Impact due to lower average speed average speed Estimated from AVL Estimated from the schedule Pax/yr Key? operating Pax operating Pax operating %capture Pax %capture 1 4,055,344 KEY 800,977 2,297, ,718 3,737,605 7,249, , % 2,575, % 23 3,714,491 KEY 912,840 2,106, ,650 3,102,271 6,404, ,556 98% 2,056,136 98% 28 4,361,924 KEY 789,009 2,387, ,324 4,269,316 7,748, , % 2,534, % 39_3 4,023,777 KEY 971,377 2,653, ,934 3,383,301 7,366,185 1,020, % 2,776, % ,008,338 KEY 1,051,494 3,426, ,059 3,524,596 8,120,895 1,235, % 4,024, % 77 2,270,881 KEY 659,692 1,057, ,756 2,115,967 4,058, ,047 89% 934,875 88% ,468 n/a 149, ,358 66, , , ,417 75% 135,163 72% ,974 n/a 87, , , , ,636 98, % 174, % ,944 n/a 131, ,635 56, , , , % 147, % 9 1,866,512 n/a 360, , , ,111,602 2,354, , % 804, % Table 14 shows that the variability impact on the operation cost on route 66 is equal to $ 118,054 /year. That doesn t mean that the service at route 66 is more reliable than the other route. To illustrate that, when the variability in running time at low traffic period is high and exceeds the variability at other periods that will create negative variability impact on the route. Thus, in most cases negative variability impact is an indicator of poor control such as in route 66. For more details check table 21 and 22 in the appendix B.

59 $/Pax 50 Then, by normalizing the impact of traffic congestion on passengers according to the estimated annual ridership, the result shows that the highest impact on passengers is found on route 66 with 1.73$/pax-trip (see figure 4) Average passenger impact of traffic congestion _ Route Figure 4 the estimated impact cost of traffic congestion on passenger Also, The total cost of traffic congestion impact on passengers and operating cost are normalized according to the annual revenue hours, which are the sum of in service time and layover time, and the results are shown in table 15 and figure 5.The highest impacts on operating cost per veh-hr is found on route 89 with value of $26. The impact on passengers per veh-hr for route 66, which is the highest, is $ Traffic congestion cost /Veh-hr _ for operating cost for passengers Figure 5 Travel congestion impact per Veh-hr

60 Using Schedule running time data 51 Table 15 Traffic congestion impact per Veh-hr Traffic congestion cost /Veh-hr ROUTE Annual revenue hours Pax/yr for operating cost for passengers $/paxtrip ,172 4,055, ,261 3,714, ,135 4,361, _3 59,902 4,023, ,655 4,008, ,752 2,270, ,575 1,181, , , ,397 1,866, Finally, using scheduled running time instead of AVL data can capture on average 103% of the real impact due to lower average speed.the range of the captured real effect percentage is between 72% and 131% of the real impact. The results shows in many cases the schedule running time can be a good estimated to the traffic congestion impact on bus due to lower average speed (See figure 6). 5,000,000 4,000,000 3,000,000 2,000,000 1,000, ,000,000 2,000,000 3,000,000 4,000,000 Using AVL Data impact on Pax Travel cost impact on Operating cost 45 degree line Figure 6 Comparing Measuring travel congestion impact on buses due to lower average speed using AVL data & Scheduled running time data

61 52 Chapter 7. Summary and Conclusions 7.1. Conclusion Even though it seems that leaving transit service without any congestion protection is the cheapest choice, transit service is incurring high loss due to traffic congestion on some routes. A case study was done in this research on 9 different bus routes. The results show that traffic congestion annual impact on operating cost for some routes may reach up to $1.7 million per route. Also, the loss on passenger may reach in some cases $6 million. Undoubtedly, passengers loss has its strong correlation with the service demand. In other words, higher travel cost leads to less likeability to maintain the same level of demand after a while. On the other hand, archived AVL data is very helpful for understanding how transit service performs. Time-point data is not all what the transit service analyzer would need to evaluate a service and suggest some improvements. Small things could be added to the system in order to make the data clear and easy to be analyzed such as a record for each different stop and a record for closing the bus door (that was discussed in depth on section 4.3) Future Research There are some gaps need to be filled in this research such as:- In this research, for lack of data availability passengers travel distance were assumed. However, researches can look at the impact on passengers travel time using O.D matrix. Considering traffic congestion impact on passenger s transfer time since transfer time is sensitive to the service reliability. Traffic congestion impact on future demand.

62 53 REFERENCES 1. Furth, P.G. "Public Transport Priority for Brussels: Lessons from Zurich, Eindhoven, and Dublin." Report to the Brussels Capital Region, Universite Libre de Bruxelles, Moynihan, D.P. The Politics and Economics of Regional Growth. The public interest, 1978,51,3-21, quoted in Fielding, Gordon J. Managing public transit strategically: a comprehensive approach to strengthening service and monitoring performance. San Francisco: Jossey-Bass Publishers, 1987, pp Massachusetts Bay Transportation Authority ScoreCard Archive. [Online] 4. Levinson H. S. Analyzing Transit Travel Time Performance. In Transportation Research Record 915, TRB, National Research Council,Washington, D.C., 1983, pp McKnight, C., Levinson, H., Ozbay, K., Kamga, C., Paaswell, R.,Impact of traffic congestion on bus travel time in northern New Jersey. Transportation Research Record 1884, 2004, Muller, Th.H.J. and P.G. Furth, Trip Time Analyzers: Key to Transit Service Quality, Transportation Research Record 1760, 2001, pp Furth, P.G., B. Hemily, T.H.J. Muller, J.G. Strathman,TCRP Report 113: Uses of Archived AVL APC Data to Improve Transit Performance and Management, Transportation Research Board, National Research Council, Washington, D.C., Cham, Laura Cecilia Understanding Bus Service Reliability: A practical framework Using AVL/APC Data. Thesis, Master of Science in Transportation. Massachusetts Institute of Technology Gerstle, David G., Understanding Bus Travel time variation using AVL data. Thesis, Master of Science in Transportation. Massachusetts Institute of Technology

63 Abkowitz, M., L. Englisher, H. Slavin, R. Waksman, and N. Wilson Transit service reliability. Report UMTA-MA , United States Department of Transportation. 11. Ceder, A. (2007). Public Transit Planning and Operation: Theory, Modeling and Practice. Elsevier, Butterworth-Heinemann, Oxford, UK. 12. Furth, P.G. and Th.H.J. Muller, "Service Reliability and Hidden Waiting Time: Insights from AVL Data," Transportation Research Record 1955, 2006, pp National Transit Database, Transit Agency Information, Boston, MA. [Online] Belenky, P., "The Value of Travel Time Savings: Departmental Guidance for Conducting Economic Evaluations," Office of the Secretary of Transportation, U.S. Department of Transportation, Washington D.C., Massachusetts Bay Transportation Authority, Bus Schedules & Maps. [Online] Furth, P.G., "Innovative Sampling Plans for Estimating Transit Passenger Miles, " Transportation Research Record 1618, 1998, pp Walker, Jarrett. on operating cost, Human Transit 2014 Web. n.d. 18. Google map throughout Massachusetts Bay Transportation Authority website.

64 55 APPENDIX A Table 16 Abbreviations Abbreviations Description Rt. Route ID hdwy Headway Class Dir Direction Period,0 0Low_Traffic Period,1 1AM Peak Period,2 2Midday Base Period,3 3Midday School Period,4 4PM Peak Period,5 5Shoulder/Eve. Period,6 6Sat_Morning Period,7 7SUN_Morning Period,8 8Afternoon_WKEND Period,9 9Evening _WKEND IB Inbound OB Outbound LH Long headway SH Short headway AVL Automated vehicle location MBTA Massachusetts Bay Transportation Authority TCI Traffic congestion impact RT Running Time nstop Average number of served stop at period p S(nStop) Standard deviation of (nstop) in period p TT Travel time H Headway S(H) Standard deviation of headway at period p DepDev percentile of departure deviation

65 56 APPENDIX B Route 23 Route 23 runs between Ashmont Station and Ruggles Station. It is one of the key bus routes in Greater Boston area. Running from Ashmont Station to Ruggles Station is the inbound direction. Figure 7 Route 23, MBTA, Boston [18]